Algorithms, data, and platforms: the diverse challenges of governing AI

ABSTRACT

Artificial intelligence (AI) poses a set of interwoven challenges. A new general purpose technology likened to steam power or electricity, AI must first be clearly defined before considering its global governance. In this context, a useful definition is technology that uses advanced computation to perform at human cognitive capacity in some task area. Like electricity, AI cannot be governed in isolation, but in the context of a broader digital technology toolbox. Establishing national and community priorities on how to reap AI’s benefits, while managing its social and economic risks, will be an evolving debate. A fundamental driver of the development and deployment of AI tools, of the algorithms and data, are the dominant Digital Platform Firms (DPFs). Unless specifically regulated, DPF's set de facto rules for use of data and algorithms. That can shift the borderline between public and private, and result in priorities that differ from those of the public sector or civil society. Governance of AI and the toolbox is a critical component of national success in the coming decades, as governments recognize opportunities and geopolitical risks posed by the suite of technologies. However, AI pries open a Pandora's box of questions that sweep across the economy and society engaging diverse communities. Rather than strive towards global agreement on a single set of market and social rules, one must consider how to pursue objectives of interoperability amongst nations with quite different political economies. Even such limited agreements are complicated following the Russian invasion of Ukraine.

KEYWORDS:

• Artificial intelligence
• AI governance
• AI regulation
• digital platform
• general purpose of technology
• global digital governance and regulation

Acknowledgements

We appreciate and wish to emphasize the involvement of Camille Carlton. We wish to thank the anonymous reviewers whose comments were most valuable. We also wish to thank Tim Buthe for his careful reading of the paper and suggestions.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 Let us not scuffle about a precise definition of ‘governing' or governance, since such precision is not needed for this discussion. Something like this will do for now:

Governance is the way rules, norms and actions are structured, sustained, regulated and held accountable. The degree of formality depends on the internal rules of a given organization and, externally, with its business partners. As such, governance may take many forms, driven by many different motivations and with many different results. For instance, a government may operate as a democracy where citizens vote on who should govern and the public good is the goal, while a non-profit organization or a corporation may be governed by a small board of directors and pursue more specific aims. (Wikipedia, accessed 14 February 2021)

2 For example, a suit under the Sherman Antitrust act of 1890 broke up Standard Oil in 1911. This kept gasoline prices low. At the same time, the regulatory setting enabled the introduction of Ford’s assembly line. Together these literally fueled Ford’s rise, selling their Model T for under $700, to become the largest and most profitable manufacturer in the world by 1912, and incidentally closing off the path to electric cars as an alternative.

3 The earlier piece was Governing AI: Understanding the Limits, Possibilities, and Risks of AI in an Era of Intelligent Tools and Systems, published by the Woodrow Wilson Center. This paper builds on this earlier publication, expanding the debate beyond the questions and issues of governance to implementing governance itself.

4 Arguably we are at a Polanyian moment with the tension between private initiative and public pushback. See Cioffi et al. (2022).

5 See, for example, McCarthy et al. (1955), Minsky (1968), Gardner (1999), McCarthy (2007), Russell and Norvig (2010), HLEG (2019). See also Büthe et al. (2022), who define AI as

systems that combine means of gathering inputs from the external environment with technologies for processing these inputs and relating them algorithmically to previously stored data, allowing such systems to carry out a wide variety of tasks with some degree of autonomy, i.e., without simultaneous, ongoing human intervention. Their capacity for learning allows AI systems to solve problems and thus support, emulate or even improve upon human decisionmaking – though (at least at this point of technological development) not holistically but only with regard to well-specified, albeit possibly quite complex tasks.

6 From vast corpora of data marked up by humans with features, identities, or labels, a system iteratively improves (‘learns') a large set of probabilities in order to predict the most likely output labels for new, never-before-encountered input data. Key to the breakthrough was using networks of simulated neurons stacked in layers, or ‘deep neural nets.' The class of techniques, called ‘deep learning,' combining these data structures and training techniques, enabled a leap in image recognition around 2011 (see, e.g., Krizhevsky et al., 2012), which put AI back on the map – some 40 years after these computational models had first been proposed (McCulloch & Pitts, 1943).

7 A system uses large scale random trial-and-error to maximize a score, a so-called objective function, such as board-position strength in chess, wind turbine efficiency, stock portfolio value, or robot motion effectiveness. When combined with deep networks, so-called deep reinforcement learning can yield systems like AlphaZero that learn, through trial-and-error, to play games like Go beyond human capacity.

8 GPT-3 is a key example of a new type of ML architecture that enables systems to ‘learn' features and structures of a class of data – in this case, written language – without human-labeling. The architecture enables so-called self-supervised learning, in which written texts are mined over and over by covering and revealing certain words, to train a language model to predict those covered words. The probabilities of grammar, usage, and style emerge in the system after it is exposed to vast corpora of books, online news sources, and publications of all sorts.

9 While this paper focuses on Narrow AI, it does not dismiss the eventual development of General AI (GAI). In fact, there is considerable effort being made to develop and understand GAI, for instance at the Center for Human Compatible Artificial Intelligence (CHAI), as well as Open AI, Deepmind, Google Brain, Facebook A.I. Research, AGI Innovations, Microsoft Research Lab, Apprente, and Kimera Systems.

10 A research emphasis of the Berkeley Roundtable on the International Economy have been the diverse ways in which DPFs are rewiring the economy and society. Kenney and Zysman (2016) describe the rise of the platform economy, whereby DPFs organize new and existing businesses into their value extraction systems. Bearson et al. (2019) present a taxonomy of platform-enabled labor and value creation, which varies substantially in terms of employment structure, compensation levels, and overall job quality. Cutolo and Kenney (2020) explore the transformative impact of DPFs on entrepreneurship, documenting the emergence of "platform-dependent entrepreneurs" whose business models depend on platforms, making them vulnerable to platform's decisions and competition. Kenney and Zysman (2019) document the ways in which DFPs finance models differ from those of the past. Kenney et al. (2019) illustrate how DFPs, once planted in traditional economic sectors, follow multiple simultaneous expansion paths to increase their presence and power across new industries.

11 George Akerlof and Robert Shiller’s extensive writings emphasize the importance of narrative in the debates amongst economists. See: Shiller (2019) and earlier Akerlof and Shiller (2009). Frederick Mayer builds related arguments in the political science literature. See Mayer (2014).

12 Cross-national differences inherently harbor the potential of serious conflict.

13 For our purposes and discussion, we set aside military and geo-strategic questions as well as the economic debates around promoting innovation.

14 See for example National Security Commission on AI chaired by Eric Schmit: Schmit (2021) or Kissinger, Schmit, and Huttenlacher (2021).

15 Here, we define AI technological leadership as distinct from access to leading edge research.

16 This has been the case historically. For instance, the technical innovation of the railroad was dependent on complementary technologies (e.g., metal for rail and engines), foundational technologies (e.g., steam), and operational training.

17 For a discussion of the impact of AI on work see. Laura Tyson and John Zysman, ‘Automation, AI, and Work’ forthcoming Daedalus Spring 2022.

18 This has been thoroughly argued by Zysman et al. (2019).

19 The implications of values being embedded into computer code have been discussed in a range of works including: Fleischmann and Wallace (2006), Groth and Nitzberg (2018), and Christian (2020).

20 ‘PizzaGate' is a widely debunked conspiracy theory which was spread through social media platforms such as 4chan, Twitter, and Reddit (Aisch et al., 2016; Wendling, 2016). PizzaGate culminated in December 2016 when an individual stormed and shot at a local D.C. pizza parlor, Comet Ping Pong, as a self-proclaimed ‘investigator' looking to ‘save the children' (Aisch et al., 2016). The theory is pointed to as a prominent example of the politically-motivated disinformation spread prior to the 2016 U.S. presidential election.

21 This estimate from this study was based on open web content only and did not include inadvertent funding on social media and video platforms (Global Disinformation Index, 2020).

22 Numerous studies point to the platform-driven disruption of retail, hospitality, and an array of industries due to digital ads. See, for example, Khan (2016), Mitchell and LaVecchia (2016), Zervas et al. (2017), Wachsmuth and Weisler (2018), Crain and Nadler (2019).

23 Kenney et al. (2019) define a mega-platform firm as a firm that operates multiple platforms across industries.

24 In November 2020, the state of California voted to approve Proposition 22, a measure allowing DPFs operating in the gig economy to continue paying workers as independent contractors. The measure, created and backed by Uber, Lyft, and DoorDash, is likely to have widespread implications for the gig economy labor laws throughout the country (Conger, 2020).

25 China’s national AI strategy is entitled ‘New Generation Artificial Intelligence Development Plan’ (新一代人工智能发展规划) and was released in 2017 (Roberts et al., 2020).

26 The White House-led strategy on AI is outlined by the National Security Commission on Artificial Intelligence (NSCAI, 2021).

27 The European Commission recently released the Digital Services Act (DSA) and Digital Markets Act (DMA), which aim to strengthen the EU single market, create a safe and trustworthy online environment, and tackle issues caused by large platforms firms (Broadbent, 2020).

Additional information

Funding

This work was supported by Ewing Marion Kauffman Foundation; German Ministry of Labour (BMAS).

Notes on contributors

Mark Nitzberg

Mark Nitzberg: Executive Director, Center for Human-Compatible Artificial Intelligence (CHAI); Director of Technology Research, Berkeley Roundtable on the International Economy (BRIE); Head of Strategic Outreach, Berkeley AI Research (BAIR).

John Zysman

John Zysman: Professor Emeritus, Political Science, University of California, Berkeley; Co-director, Berkeley Roundtable on the International Economy (BRIE); Convenor of the WITS Project through BRIE and the Center for Information Technology Research in the Interest of Society (CITRIS).